How much of GitHub's most-starred AI tooling is written by coding agents?
AI authorship, churn and self-published results in today's most-starred AI-tooling repositories
Five Observations
AI Authorship Is Common
AI-Signed Share of Commits, Last 180 Days
One bar per repository • colour = AI evidence class
Older Projects Are Switching
🔍 Unsigned Does Not Mean "No AI"
Lower bound| Repository | Signed | Agent configuration committed |
|---|
AI Share Does Not Track Popularity
Stars vs AI-Signed Share (Last 180 Days)
Log scale on stars • click a dot to open it in the explorer
More Fixing, Not More Throwing Away
AI-Era Cohort vs Famous Pre-Copilot Projects, at the Same Age
Each dot is a repository checked out at that age • the bar is the median • hover for names
| Median | Cohort | Baseline | P(cohort > baseline) | p-value |
|---|
What They Publish About Their Own Results
Strongest Published Evidence per Repository
How each repository backs the numbers it publishes about itself
🧪 Outcome benchmarks with a harness and committed results
📉 Results That Did Not Go Their Way
Publishing these is unusual, and is itself evidence of an evaluation habit.
📚 All evidence records
Scorecard
Click a row to open it in the explorer. AI-signed % is a lower bound; fix % is keyword-based.
Repository Explorer
AI signals, most-changed files and the latest authored commits of any audited repository.
Most-Changed Files (Hotspots)
In-scope source and Markdown files, by number of commits
| File Path | Commits | Added | Deleted | Churn |
|---|
Latest Authored Commits
With the AI tool named by each commit's signature
FAQ & Related Research
The questions a careful reader will ask, answered with the limits stated, followed by the research this study builds on. Most published work measures the average effect of AI tools across many projects and finds rising churn and quality costs; this study looks at the top of the popularity distribution instead. The two are compatible.
📖 Related research
Downloads & Citation
The scorecard, every headline number with the evidence behind it, the written report and a citation. The complete dataset (per-repository metrics, age-matched snapshots and every evidence record) is on GitHub under CC BY 4.0.
Full dataset on GitHub📦 Download Hub
Scorecard Spreadsheet
One row per repository: stars, commits, AI evidence class, AI-signed share (all-time and last 180 days), fix share, churn, survival and published-evidence type.
Findings & Evidence Database
Every headline number with its statistics, the baseline snapshots, and the per-repository evidence records with file and line references.
Written Report
The same findings as text, generated from the data: method, limits, related work, evidence links and the full scorecard.